叶绿素是植被光合作用的重要色素, 传统实验室方法测定叶绿素含量需破坏性取样且操作复杂。 通过构建高精度SPAD光谱估算模型, 可以实现对水稻叶片叶绿素含量的实时无损监测。 以黑龙江省不同施氮水平下水稻为研究对象, 采用SVC HR768i型光谱辐射仪共获取移栽后、分蘖期、拔节期、孕穗期、抽穗期共五个关键时期水稻叶片反射光谱数据。 光谱探测范围350~2 500 nm。 利用自带光源型手持叶片光谱探测器直接测定叶片光谱, 光源为内置卤素灯。 采用SPAD-502型手持式叶绿素仪同步测定水稻叶片的SPAD值。 叶片水分是植物光合作用的基本原料, 也间接影响着叶绿素含量。 叶片含水量降低则会影响植物正常的光合作用, 导致其叶绿素含量随之降低。 因此将叶绿素敏感波段与水分吸收范围结合作为SPAD估算的输入量。 随机森林模型是一个基于多个分类树的算法。 算法在采样的过程中包括两个完全随机的过程, 一是有放回抽样, 可能会得到重复的样本, 二是选取自变量是随机的。 因此本文对叶片光谱反射率进行去包络线(CR)处理, 综合考虑可见光近红外波段提取水稻叶片反射光谱特征参数和植被指数, 综合分析光谱指标与SPAD相关关系, 采用随机森林算法构建不同输入量的SPAD高光谱估算模型。 结果表明: (1)水稻叶片SPAD与光谱反射率的相关系数在叶绿素敏感波段红波段范围(600～690 nm)、红边范围(720~760 nm)、水分吸收波段范围(1 400~1 490和1 900~1 980 nm)均为0.75以上; (2)在光谱参数与SPAD 的相关分析中, NDVI, DP2与水稻叶片SPAD值相关性最好, 相关系数为0.811和0.808; (3)以结合水分光谱信息后的CR(V1, V2, V3, V4)为自变量所建立的随机森林模型精度最高, R2为0.715, RMSE为2.646, 可作为水稻叶片叶绿素预测模型。 研究结果揭示了不同品种水稻的光谱响应机制, 提供了水稻叶片SPAD值高精度反演的技术方法, 为监测与调控东北地区水稻正常生育进程提供技术支持。

Chlorophyll is an important pigment in vegetation photosynthesis, and the traditional laboratory method needs destructive sampling and complex operation. By constructing a high-precision SPAD spectral estimation model, the real-time non-destructive monitoring of chlorophyll content in rice leaves can be realized. In this paper, the data of five key stages of transplanting, tillering stage, jointing stage, booting stage and heading stage were obtained from rice under different nitrogen levels in Heilongjiang Province. The reflectance spectrum data of rice leaves were measured by SVC HR768i spectral radiometer with a spectral detection range of 350~2 500 nm. The spectrum of the blade was measured directly by the handheld blade spectrum detector with its own light source, which was built-in halogen lamp. The SPAD value of rice leaves was measured synchronously by SPAD-502 hand-held chlorophyll meter. Leaf water is the basic raw material of plant photosynthesis, and the decrease of leaf water content will affect the normal photosynthesis of plant, resulting in the decrease of chlorophyll content and the indirect effect of water content on chlorophyll content. Therefore, the chlorophyll sensitive band and the range of water absorption are combined as the input of SPAD. The Random Forest model is an algorithm based on multiple classification trees. In the process of sampling, the algorithm includes two completely random processes, of which one is that the sampling process is carried out with a return sampling, and the other is that the sample may be repeated, and the other is random when we select the independent variables. In this paper, the spectral reflectance of rice leaves is extracted by continuum removal (CR), and the characteristic parameters of reflectance spectrum and vegetation index of rice leaves are extracted by taking into account the visible and near infrared bands. The correlation between spectral indices and SPAD was analyzed and the SPAD hyperspectral estimation model with different inputs was constructed by the Random Forests. Results are: (1) The correlation coefficient between SPAD and spectral reflectance of rice leaves was above 0.75 in the range of chlorophyll sensitive band (600～690 nm), red edge region (720~760 nm) and water absorption band (1 400~1 490, 1 900～1 980 nm). (2) In the correlation analysis between spectral parameters and SPAD, the correlation between, NDVI, DP2 and SPAD value of rice leaves was the best, and the correlation coefficients were 0.811 and 0.808; (3) The Random Forests model with CR(V1, V2, V3, V4) combined with water spectral information had the highest accuracy and R2 was 0.715, RMSE was 2.646, which could be used as a chlorophyll prediction model for rice leaves. The results revealed the spectral response mechanism of different varieties of rice, provided a high precision inversion method of SPAD values of rice leaves, and provided technical support for monitoring and regulating the normal growth process of rice in Northeast China.